Systems capable of performing speech recognition are well known in the prior art. These are systems which respond to a spoken word by producing the textual spelling, or some other symbolic output, associated with that word. Commonly, speech recognition systems operate in the following manner. First, they receive from a microphone an electrical representation of the acoustic signal generated by the utterance of the word to be recognized. In FIG. 1 a simplified representation of such an acoustic signal 100 is shown in the form of a spectrogram, which plots frequency along the vertical axis, time along the horizontal axis, and the intensity of the sound at any given frequency and time by degree of darkness. Such systems normally receive such signals as an analog waveform generated by a microphone, which corresponds to the variations in air pressure over time associated with the sound of a spoken word. As they receive such signals they perform an analog to digital conversion, which converts the amplitude of the acoustic signal into a corresponding digital value at each of a succession of evenly spaced points in time. Commonly, such sampling is performed between 6,000 to 16,000 times per second for speech recognition. Once a digital representation of the amplitude waveform is obtained, digital signal processing is performed upon that digital waveform. For example, in prior art DragonDictate speech recognition systems digital signal processing is used to take an FFT, or fast Fourier transform, of the signal. This produces a digitized spectrogram representation 102 of the signal shown in FIG. 2. This spectrogram provides a vector, or frame, 104 for each 50th of a second. Each such frame is an ordered succession of values which represents the intensities at each of seven frequency ranges for each such 50th of a second. Although not shown in FIG. 1 or FIG. 2, the vector 104 also includes an energy term which represents the overall sound energy for each fiftieth of a second, and eight cepstral parameters. These cepstral parameters provide frequency-related information for each fiftieth of a second which focuses on that part of the total speech signal which is generated by a user's vocal tract, and, thus, which is particularly relevant in speech recognition.
Once a series 102 of frames 104 is produced for an utterance, as is shown in FIG. 2, that series 102, which we call a token, is matched against each of a plurality of word models 108 to find which of them it most closely matches. As is shown in FIG. 2, when this matching is performed, a process known as time aligning seeks to stretch or compress successive portions of the word model 108 as it is fitted against the token model 102 to achieve the best match. On FIG. 2, this is shown, for example, by the mapping of the two token vectors 104A against the single word model vector 109A, and the mapping of the three vectors 104B against the single model vector 109B. When this comparison is done, silence models 110 and 112, respectively, are put at the beginning and end of each word model. This is done because the utterance to be recognized will normally be preceded and followed by silence in a discreet utterance recognizer, in which words to be recognized are to be spoken separately.
FIG. 3 schematically represents the recognition process, in which the process of time aligning shown in FIG. 2 is performed between the utterance model 102 and each of the plurality of word models labeled 108A through 108N. The circles with loop arrows on top of them shown in FIG. 3 correspond to the model vectors 109 shown in FIG. 2, which also have looped arrows on top of them. The looped arrow represents the fact that when the time aligning occurs a given frame, or vector, in the word model can be mapped against one or more vectors of the token. A score is given to each of the mappings, indicating how similar the vectors of the token are to those of each of the word models they are mapped against. The word whose word model has the best score is normally considered the recognized word.
The above description of the basic operation of a speech recognition system is a highly simplified one. Much more detailed descriptions of such systems is given in U.S. Pat. Nos. 4,783,803, issued to James K. Baker et al. on Nov. 8, 1988, and entitled "Speech Recognition Apparatus And Method"; 4,903,305, issued to Laurence Gillick et al. on Feb. 20, 1990, and entitled "Method for Representing Word Models For Use In Speech Recognition"; 4,866,778, issued to James K. Baker on Sep. 12, 1989, and entitled "Interactive Speech Recognition Apparatus", and 5,027,406, issued to Jed Roberts et al. on Jun. 25, 1991, and entitled "Method For Interactive Speech Recognition And Training". The patents have all been assigned to the assignee of the present invention, and they are all hereby incorporated by reference herein.